MultiScale-Hybrid-e5-humdisp-adv detector submission

[Vigora]

This PR adds the MultiScale-Hybrid-e5-humdisp-adv detector submission to the RAID leaderboard.

It is identical to MultiScale-Hybrid-e5-humdisp (opened separately) in its features, its projection direction and its classifier settings. The single difference is the training set: 25,876 attacked documents from the RAID training split are added to the unattacked ones, each keeping the label of the text it was derived from.

The two submissions are made together so that the contribution of adversarial training can be measured on the hidden test split rather than estimated internally. Our own held-out estimate of the difference is +0.7 points of TPR at FPR=1%, which is below the resolution of our internal evaluation.

The detector was trained on the RAID training split and should be categorized in the "trained on RAID" section. Method details will be described in a paper currently in preparation.

Submission files: metadata.json, predictions.json

[github-actions]

Eval run succeeded! Link to run: link

Here are the results of the submission(s):

MultiScale-Hybrid-e5-humdisp-adv

Release date: 2026-09-04

I've committed detailed results of this detector's performance on the test set to this PR.

On the RAID dataset as a whole (aggregated across all generation models, domains, decoding strategies, repetition penalties, and adversarial attacks), it achieved an AUROC of 98.88 and a TPR of 96.19% at FPR=5% and 88.76% at FPR=1%.
Without adversarial attacks, it achieved AUROC of 99.43 and a TPR of 97.98% at FPR=5% and 92.90% at FPR=1%.

If all looks well, a maintainer will come by soon to merge this PR and your entry/entries will appear on the leaderboard. If you need to make any changes, feel free to push new commits to this PR. Thanks for submitting to RAID!